chptToBin Namespace Reference

Functions
def	chpt_to_dict_arrays_simple (file_name)
def	chpt_to_dict_arrays ()
def	concatenate_layers (params)
def	convert_rnn_kernel (weights, dimensions, is_decoder_rnn=False)
def	convert_rnn_bias (weights, dimensions, forget_bias=1.0)
def	convert_weigts (dimensions, data, forget_bias=1.0)
def	save_layer_weights (data, list_keys, dims, footer_string, file_name)
def	main (_)

Function Documentation

chpt_to_dict_arrays_simple()

def chptToBin.chpt_to_dict_arrays_simple

(

file_name

)

    Convert a checkpoint into into a dictionary of numpy arrays
    for later use in TensorRT NMT sample.

chpt_to_dict_arrays()

def chptToBin.chpt_to_dict_arrays

(

)

    Convert a checkpoint into a dictionary of numpy arrays
    for later use in TensorRT NMT sample.
    git clone https://github.com/tensorflow/nmt.git

Here is the caller graph for this function:

concatenate_layers()

def chptToBin.concatenate_layers

(

params

)

Concatenate weights from multiple layers

Here is the caller graph for this function:

convert_rnn_kernel()

def chptToBin.convert_rnn_kernel	(		weights,
			dimensions,
			is_decoder_rnn = False
	)

In place. weights conversion
TensorFlow weight parameters for BasicLSTMCell
are formatted as:
Each [WR][icfo] is hiddenSize sequential elements.
CellN  Row 0: WiT, WcT, WfT, WoT
CellN  Row 1: WiT, WcT, WfT, WoT
...
CellN RowM-1: WiT, WcT, WfT, WoT
CellN RowM+0: RiT, RcT, RfT, RoT
CellN RowM+1: RiT, RcT, RfT, RoT
...
CellNRow(M+P)-1: RiT, RcT, RfT, RoT
M - data size
P - projection size
TensorRT expects the format to laid out in memory:
CellN: Wf, Wi, Wc, Wo, Rf, Ri, Rc, Ro

For the purpose of implementing LSTMP all W and R weights become weights from W
CellN: Wf, Rf, Wi, Ri, Wc, Rc, Wo, Ro, Empty states

Update: alternative notation
Tensorflow documents gates' order in e.g.
https:github.com/tensorflow/tensorflow/blob/r1.4/tensorflow/python/ops/rnn_cell_impl.py:439
TF: i = input_gate, j = new_input (cell gate), f = forget_gate, o = output_gate - ijfo
Need to convert 'ijfo' to 'fijo'

Here is the caller graph for this function:

convert_rnn_bias()

def chptToBin.convert_rnn_bias	(		weights,
			dimensions,
			forget_bias = 1.0
	)

TensorFlow bias parameters for BasicLSTMCell
are formatted as:
CellN: Bi, Bc, Bf, Bo

TensorRT expects the format to be:
CellN: Wf, Wi, Wc, Wo, Rf, Ri, Rc, Ro

Since Tensorflow already combines U and W,
we double the size and set all of U to zero.

Here is the caller graph for this function:

convert_weigts()

def chptToBin.convert_weigts	(		dimensions,
			data,
			forget_bias = 1.0
	)

Convert weights from Tensorflow to TensorRT format

Here is the call graph for this function:

Here is the caller graph for this function:

save_layer_weights()

def chptToBin.save_layer_weights	(		data,
			list_keys,
			dims,
			footer_string,
			file_name
	)

    data          - dictionary with string names as keys and
                    numpy weights as values
    list_keys     - list of dictionary keys to save
    dims          - list of int values relevant to the layer
                    e.g. tensor dimensions sufficient to extract all the tensors
    footer_string - marker placed at the end of file

    file format: data -> meta_data -> footer

Here is the caller graph for this function:

main()

def chptToBin.main

(

_

)

Here is the call graph for this function: